TA session# 12
Jun Sakamoto
July 20,2016
Contents
1
Endogenous Problem 1
2
Method of moment 3
3
IV; Instrumental variable methods 3
4
2SLS; 2 steps least square 4
5
Examples of IV and 2SLS 7
1 Endogenous Problem
If relation of variables and error are not orthgonal, then a crirical problems are happen in analysis. 1. Model
We consider models of demand function and supply function as below.
PD= αD+ βDQ+ ǫD PS = αS+ βSQ+ ǫS
E[ǫD] = E[ǫS] = 0, V ar[ǫD] = σd2, V ar[ǫS] = σ2s, Cov[ǫD, ǫS] = 0 We can observe prices only at equiribrium. So sample is PD= PS = P . Thus Q is
Q=(α
D−αS)+(ǫD−ǫS) βS−βD
We consider the covariance of Q and error.
Cov[Q, ǫS] = E[(α
D−αS)+(ǫD−ǫS) βS−βD ǫ
S] = σ2s
βS−βD
We can understand that Q and error are not orthgonal. Then, what happen OLS estimator? V ar[Q] and Cov[P, Q] are
V ar[Q] = E[((α
D−αS)+(ǫD−ǫS) βS−βD )
2]
=(βσS2d−β+σD2s)2
Cov[P, Q] = β
SσD2+βDσS2
(βS−βD)2
βsis Cov[P,Q]V ar[Q] . So we can get
βs= β
Sσ2d+βDσ2s
σd2+σ2s
= θβS+ (1 − θ)βD θ=σ2σd2
d+σ2s
So OLS estimator of β is weighted average of βS and βD. Thus β has neither unbiasedness nor consistency without θ = 0 or θ = 1.
Next we consider the problem of estimated error.
Y = Xβ + u Z = X + e
We can not observe X and Z is obtained. (For simplicity, X and Z are vector, and β is scalar)Then we estimate a model as below.
Y = Zβ + r Then, this model can be transform as below.
Y = (Z − e)β + r Y = Zβ + (u − eβ) So error is r = (u − eβ). Thus E[Zu] is
E[Xu] = E[(X + e)(u − eβ)] = −βV [e] 6= 0
Thus Z is endogenous variable. Then, OLSE is not unbiaseed and consistent.Because normal OLSE is
E[ ˆβ] = β + E[(X′X)−1X′u] E[(X′X)−1X′u] is not usually 0. Next we check consistency.
βˆ− β = (1n ixix′i)−1 1n ixiui
= Q−1xxQxu
Then by the LLN
1
n ixiui−→pE[xiui] 6= 0 Thus
plimQ−1xxQxu6= 0
Hence, β is not consistent.
2 Method of moment
Moment condition is satisfied as below
E[g(X; θ)] = 0
where g() is function vector with dimensional k. X = X1, X2, ..., Xn is sample. Then moment estimator θ is defined as the paramater which satisfies below equation.
1
nΣig(X; θ) = 0.
3 IV; Instrumental variable methods
Some regression model has endgenous problem. Let z correlate with explanatory variable x and orthogonal with error u. Then we can estimate consistent paramator by IV.(z is called as instrumental variable.) We consider a regression model as below.
y= Xβ + u E[zixi] = Mzxis regular.
E[ziui] = 0
ui, xi, zi has forth order moment. IV estimator ˆβIV is obtained by method of moment.
E[ziui] = E[zi(yi− x′iβ)] = 0 is assumed.(moment condition) Moment estimator is
1
nΣi[zi(yi− x′iβ)] = 0
βˆIV = [Σix′i]−1Σizixi
= (Z′X)−1Z′Y Consistency
βˆIV = (Z′X)−1Z′Y (Z′X)−1Z′(Xβ + u)
= β + (Z′X)−1Z′u
= β + (1nΣizixi′)−1 1nΣiziui
By assumption and LLN,
1
nΣizix′i→pE[zixi] = Mzx<∞
1
nΣiziui→pE[ziui] = 0 So we can get,
βˆIV →pβ
Asymptotic normality
√n( ˆβ
IV − β) =√n(Z′X)−1Z′u
= (n1Σizixi′)−1 1nΣiziui
We use E[ziui] = 0 and V [ziui] = E[u2izizi′] = S < ∞ from assumption. Then by CLT and normality
√n( ˆ
βIV − β) →d Mxx−1N(0, S)
= N (0, Mxx−1SMxx−1)
4 2SLS; 2 steps least square
We consider the case of multiple endogenous variables, exogenous variable and IV.
y= X1β1+ X2β2+ u
where X1:matrix of endogenous variables, X2:exogenous variable and Z:IV. l is a number of IV and k is a number of endogenous variables. We can consider 3 cases.
l > k:over identified l= k:just identified
Now we consider the case of over identified. Thus if IV are more than endogenous variables, Then we use 2SLS. First we estimate a below equation.
X1= Z1δ1+ X2δ2+ v
We generate predict values ˆX1 by parameters of above equation. This equation is called as reduced form. We use this values in second step regression as below.
y= ˆX1β1+ X2β2+ u
Estimator of above equation is same as IV estimator which employ ˆX1and X2as IV. Next we define below.
X:A matrix of all variables.((endogenous variables)+(exogenous variables)) Z:A matrix of all exogenous variables.
Xˆ:A matrix of ( ˆX1, X2) Predict value ˆX is obtained by,
Xˆ = Z(Z′Z)−1Z′X Because coefficients of reduced form are written as below.
δ= (Z′Z)−1Z′X Thus ˆβ2SLS is
βˆ2SLS = ( ˆX′Xˆ)−1Xˆ′y
= ((Z(Z′Z)−1Z′X)′Z(Z′Z)−1Z′X)−1(Z(Z′Z)−1Z′X)−1y
= (X′Z(Z′Z)−1Z′Z(Z′Z)−1Z′X)−1X′Z(Z′Z)−1Z′y
= (X′Z(Z′Z)−1Z′X)−1X′Z(Z′Z)−1Z′y Next we define some assumption to confirm asymptotic properties of 2SLS.
E[zix′i] = Mzxis regular. E[z′z] = Mzz is regular.
E[ziui] = 0
All variables have forth order moment.
Consistency
βˆ2SLS = (X′Z(Z′Z)−1Z′X)−1X′Z(Z′Z)−1Z′y
= (X′Z(Z′Z)−1Z′X)−1X′Z(Z′Z)−1Z′(Xβ + u)
= β + (X′Z(Z′Z)−1Z′X)−1X′Z(Z′Z)−1Z′u
= β + (n1Σixiz′i(n1Σizizi′)−1 1nΣizixi′)−1 1nΣixizi′(1nΣiziz′i)−1 1nΣiziui
By LLN and assumption,
1
nΣixizi′→pE[xiz′i] = Mzx′ <∞
1
nΣiziz′i→pE[ziz′i] = Mzz<∞
1
nΣizix′i→pE[zix′i] = Mzx<∞
1
nΣiziui→pE[ziui] = 0 By the continuous mapping theorem,
(n1Σixiz′i(n1Σizizi′)−1 1nΣizix′i)−1→p(Mzx′ Mzz−1Mzx) < ∞
(n1Σizizi′)−1→pE[zizi′] = Mzz−1<∞ So, we can get,
βˆ2SLS→pβ
Asymptotic normarity
√n( ˆ
β2SLS− β) = (1nΣixizi′(1nΣizizi′)−1 1nΣizixi′)−1 1nΣixizi′(n1Σiziz′i)−1 1√nΣiziui
By the LLN,
(1nΣixizi′(1nΣiziz′i)−1 1nΣizix′i)−1 →p(Mzx′ Mzz−1Mzx)−1
1
nΣizix′i →p Mzx 1
nΣiziz
i′ →p E[zizi′] = Mzz−1
E[ziui] = 0, E[u2izizi′] < inf ty and by the CLT,
√1nΣiziui→dN(0, E[u2izizi′])
= N (0, Ω) Thus, by the CLT,
√n( ˆβ
2SLS− β) →d(Mzx′ Mzz−1Mzx)−1MzxMzz−1N(0, Ω)
= N (0, (M′ M−1M )−1M M−1ΩM−1M (M′ M−1M )−1)
5 Examples of IV and 2SLS
We consider a under example.
*IV example*
y= βX + u X = δz + γu + e
Parameters are set as β = 1.5, δ = 2, γ = 0.8. Also z ∼ U(−5, 5), u ∼ N(0, 1), e ∼ N(0, 1) Experiment steps
1.We generate samples z, u and e. 2.y and X made from above model. 3.We estimate ˆβIV and ˆβOLS.
4. ˆβIV and ˆβOLS are obtained by repeating above steps 10000 times.
*2SLS example*
G.D.Pietro(2012)”Does studying abroad cause international labor mobility? Evidence from Italy” Economic Letters 117 pp632-pp635
This paper confirm relationship of studying abroad and international labor mobility.Several studies(Parey and Waldinger etc)shows that there is a positive relationship between international student mobility and the desicion to work abroad following graduation.
However that studies faced omitted variable bias. Thus, this paper use 2SLS method.
Model
The following baseline specification can be used to investigate the effect of studying abroad on the decision to work in a foreign country.
Yijk= β0+ β1studyabroadijk+ β2Xijk+ β3Dj+ β4Uk+ ǫijk
Y:takes the 1 if he/she works abroad and 0 other wise. studyabroad:takes the 1 if he/she study abroad and 0 otherwise.
X, D, U:control variables. And reduced form is
studyabroadijk= α0+ α1ex− studyabroadijk+ α2Xijk+ α3Dj+ α4Uk+ ǫijk ex− studyabroad is a measure of students’ exprosure to study abroad programs.
Figure 1:
Figure 2: